acceptodds
Withdrawn from ICLR 2027

NeuronDiscover: Agent-in-Twin for Mechanistic Discovery in Neuronal Microenvironments with World Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Mechanistic discovery in neuronal microenvironments requires identifying interventions and observations that discriminate between competing biological explanations. We formulate this problem as joint inference over mechanistic hypotheses and digital twin states, and propose **NeuronDiscover**, an *Agent-in-Twin* framework that integrates scientific reasoning, intervention planning, and evidence acquisition within a mechanism-grounded *World Action Model (WAM)*. NeuronDiscover represents biological knowledge through a structured *Mechanism–Intervention–Observation–Outcome (MIOY)* graph, where experimental evidence updates mechanistic beliefs and guides targeted twin refinement. The agent generates executable intervention programs with explicit verification criteria and adapts its discovery strategy according to uncertainty and observation constraints. We evaluate NeuronDiscover on controlled brain-fluid transport environments with independently verified reference simulations and donor-disjoint cortical neuron electrophysiology recordings. Experiments demonstrate that Agent-in-Twin enables more accurate mechanism identification, selective twin refinement, and efficient intervention prioritization under limited observation budgets. This work introduces an adaptive scientific agent paradigm in which digital twins evolve as mechanistic models through evidence-driven interaction, enabling autonomous discovery in complex neuronal systems. The code is available at https://anonymous.4open.science/r/NeuronDiscover.

Primary area:
applications to neuroscience & cognitive science
Keywords:
scientific agents, scientific digital twins, world action models, Agent-in-Twin, Twin confounding, bidirectional scientific discovery, mechanistic discovery, neuronal microenvironments

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qyH5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

NeuronDiscover combines a mechanistic simulator, a learned world model, adaptive experiment selection, and a revisable hypothesis graph. It chooses interventions and measurements to distinguish biological mechanisms from simulator or sensor errors, then updates claims using independent evidence. Also, evaluation covers simulated brain-fluid transport and neuronal recordings from held-out donors

Critical strengths
  1. This paper addressed an important scientific inference problem -- explicitly separating mechanism uncertainty from simulator discrepancy helps prevent model errors from being mistaken for biological explanations. Also, its prospective testing, scoped claims, and preserved counterexamples make the discovery process more auditable.
  2. The paper demonstrates that there are substantial gains with substantial experimental controls. The method reports 4 resolved relations per world, compared with 3.4 for the matched external agent, while reducing false support from 8% to 5%. The paired improvement is +0.60 [95% CI: 0.41-0.79] (statistically significant) across 32 independent sources. Also, the component ablations and deterministic controls help establish where these gains arise (Tables 2-4).
Critical weaknesses
  1. The paper's claimed certification savings rely on an insufficiently justified calibration assumption. For example, equation 50 uses reference samples from each program’s outcome distribution, whereas Table 23 assumes that a single shared reference batch can calibrate all programs. Sharing event definitions does not establish that reuse is valid, and this matters because the directly demonstrated comparison favors direct reference certification: 8/8 programs certified versus 5/8, using the same 65,536 reference executions (Tables 21-22).
  2. Evidence for broader mechanistic discovery remains limited. Transport discovery operates within a supplied mechanistic ontology and executable model family. The real-recording tasks recover input resistance, firing gain, and adaptation, without establishing underlying biological mechanisms. These are useful demonstrations, but leave generalization beyond the designed hypothesis families uncertain.
Questions for authors
  1. What sampling or transfer argument makes the shared calibration batch in Table 23 valid for every program-specific discrepancy bound? Please account for all reference executions and explain whether the direct-certification baseline could reuse the same evidence.
  2. In omitted-relation tasks, what is withheld from the ontology and executable mechanism templates, versus merely omitted from the initial graph? Please provide a representative recovered relation and explain why the matched deterministic controller cannot generate it.
  3. Can you please provide a working link to your code?
Reviewer_hoEBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents NeuronDiscover, a framework for mechanistic discovery and lab in the loop agentic science. The authors benchmark this framework on two problems: A dynamical system simulating fluid transport, and a neuronal benchmark.

Critical strengths

The problem the authors are trying to solve is an interesting and difficult one. The authors do disclose some of the limitations of their work.

Critical weaknesses

I have a lot of concerns about this paper. Which I list below:

Clarity

  • The authors do not clearly describe the problem they are trying to solve. The abstract and introduction talk about lab in the loop, and about agent-twin loops; and figure 1 talks about “autonomous discovery in complex neuronal systems” but that is not what the rest of the paper does at all. Instead, the paper benchmarks on a simulator where the “real-world” is just a finer simulation; and on a setting where the “real world” in an archive of past recordings. None of these settings even begin to simulate the complexities of real world science, and the difficulties of simulating it. In fact, using AI to simulate coarse grained simulations from finer ones is a field with extensive work, none of which the authors cite (more on that below).
  • Undefined symbols: In section 3, the authors introduce equations containing symbols that have never been introduced before (some examples are MtM_t in equation 1, Res/False in equation 2, ℓV\ell_V and VV on equation 8, but there are many others.
  • Similarly to the point above, there are many undefined terms that are introduced, seemingly key to the paper, without either a citation or an explanation such as "GlymphTwin" and "I-MDD-WM”. Some of these might be introduced in the (45 page!) appendix, but in many cases the main text does not even point to a place in the appendix where this information can be found.
  • Example 17 is another example of the above: It describes the entire PDE that the algorithm is modeled, but it has been relegated to an appendix, and the main text just mentions a “concentration, pressure, flow… “ out of the blue.
  • The figures bring very little clarity. They are extremely crowded, with a lot of text, and do not at all help convey the point they are trying to convey

Conceptual

The main issue here is that the theory is mostly basic results presented as novel. Examples are:

  • Prop 3 says conditionals of one joint are compatible.
  • Lemma 1 says clamped values stay clamped.
  • Lemma 2 is the triangle inequality

My other conceptual issues is, why do we need a learned diffusion model for a 17-cell PDE? It seems like overkill

Finally on this point, the experiments are almost trivial, and do not at all reflect on the difficulty of the problem discussed in the abstract and introduction.

Missing citations

  • The problem the authors call “twin confounding” is basically the Kennedy-O’Hagan problem. This should be mentioned with the appropriate references (Kennedy & O'Hagan (2001), Brynjarsdóttir & O'Hagan (2014))
  • There is also extensive literature on using AI to go from fine to coarse simulations, e.g. Um et al. (2020) Kochkov et al. (2021)
  • There is also extensive literature on choosing experiments to discriminate between rival mechanisms (such as Box & Hill (1967) Atkinson & Fedorov (1975))

Typos/Minor issues

  • Section 5 says "We ask four questions" and then lists three.
  • Figure 2 lists NEURON, Brian2 and BluePyOpt as tools, but nothing in the experiments mentions using them.

Conclusions

The paper is well below the bar for the venue, and I recommend its rejection

Questions for authors
  • Could the authors state the PDE, the observation model, and the candidate explanations in the main text?
  • Could the authors make sure every symbol in the equations is previously introduced in the main text?
  • Which components of Figure 1 are implemented in this work, and which are simulated or left for future work? Could this be clarified in the text and/or caption?
  • Which theoretical results do the authors consider novel contributions?
  • Why is a learned diffusion model needed for a 17-cell PDE?
Reviewer_GEaPOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses mechanistic discovery when the simulator used to plan experiments (the "digital twin") may itself be wrong. It defines twin confounding as the case where the available experiments cannot separate a difference in biological mechanism from a compensating error of the twin, and proposes NeuronDiscover, an "Agent-in-Twin" system with three parts: (i) a World Action Model (WAM), a masked discrete-diffusion model over tokenized finite-volume states, coupled to a coarse physics solver and a learned residual, that answers forward, inverse, and observation-design queries; (ii) an agent (an LLM proposal interface plus a numerical controller) that selects interventions and observations, first by raising a local Fisher-information floor and then by expected information gain over the joint mechanism and discrepancy belief; and (iii) a typed Mechanism-Intervention-Observation-Outcome (MIOY) graph whose relations are tested only on outcomes released after the selection is frozen, using chi-square set inversion, error spending across relation births, and certified response enclosures based on a posteriori finite-element bounds. Experiments use synthetic brain-fluid transport worlds (GlymphTwin; 32 generator families; a 17-cell operational mesh against a 136-cell reference) and donor-disjoint Allen Cell Types current-clamp recordings with three response predicates. Relative to a fixed hypothesis graph, graph revision raises certified resolutions from 3.2 to 4.0 per world and lowers false support from 7.0% to 5.0%; joint mechanism-discrepancy EIG beats plug-in and nested-filter EIG; and the mechanism-grounded WAM gives better discovery than an FNO that has lower rollout error.

Critical strengths

S1. The question is important. Discovery systems that plan with a simulator need to tell "the mechanism is wrong" apart from "the simulator is wrong", and most current AI-scientist evaluations do not measure this. Scoring by resolved relations, false support, scope accuracy, and abstention instead of rollout error is the right direction, and the FNO versus hybrid result (better NRMSE, worse discovery) is a useful illustration of why.

S2. The statistical design is careful for an agent paper. Selections are frozen before outcomes are released, error is spent across adaptively born relations, set inversion keeps bound directions correct under directed rounding, and comparisons are paired at the level of independent source units, with exact permutation tests where the number of units is small.

S3. The MCP tool schema and prompts (Appendix D.7) and the replayed discovery trace (Appendix K.4) make the system auditable, and the worked numbers can be checked by hand.

Critical weaknesses

W1. The core concept overlaps with the classical computer-model calibration literature, which is not cited; the only discrepancy-learning method the paper cites is the one used as its discrepancy-aware baseline (Table 2). Several theoretical results are standard tools (Proposition 1 is local identifiability through a positive-definite Fisher information; Proposition 2 is a total-variation bound; Theorem 1 is Hoeffding plus a union bound; Propositions 6 and 7 are super-uniformity after freezing and error spending). They are fine as tools, but the paper should say that they are standard results applied to this setting.

W2. The evidence does not support "mechanistic discovery in neuronal microenvironments". (a) In GlymphTwin, reference worlds are generated from the declared model of Eq. (17) on a 136-cell mesh and differ from the 17-cell twin by declared dynamics, boundary, and sensor differences (Appendix G.3), so the true explanation lies inside the declared envelope by construction, except in the stress test of Appendix I.4. Discovery is therefore selection among pre-specified compositions plus scope refinement, and the "twin discrepancy" consists of discretization error and declared discrepancy coordinates. The system is never run on real tracer data, although public human data have been used to estimate diffusion versus convection in exactly this setting.

W3. The twin is not refined. The abstract claims "selective twin refinement" and twins that "evolve as mechanistic models through evidence-driven interaction", but the WAM is trained offline once and kept fixed during discovery (the round prompt in Appendix D.7 says "Keep reset state and Twin version fixed"; see also Appendices J.4 and L). What changes is the belief over discrepancy coordinates and the MIOY graph. Please either show the twin (WAM or residual) being updated from acquired evidence and the effect of doing so, or revise these claims.

W4. The agent adds little relative to the machinery, and the most natural simple baselines are missing. The deterministic controller reaches 3.8 versus 4.0 resolutions (Table 2), and Table 11 shows that the aggregate +0.20 comes from one task family with eight sources (Holm p = 0.031, which the paper notes is the smallest attainable value). Most of the gain over baselines comes from the numerical controller. I would make the controller the headline and justify the cost of the LLM.

Questions for authors

Q1. If the operational twin used the 136-cell mesh, does twin confounding largely disappear? How do the results change with operational mesh resolution?

Q2. What are the Table 2 endpoints for physics-only, for a KOH-style GP discrepancy model, and for NPE with the solver, under the same controller and budget?

Q3. How many relations does the agent propose that match no reference relation, and how are they scored?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.